anuragsahu/stereo-dense-reconstruction

This repo contains the code for Stereo Dense Reconstruction, Given the Stereo image pairs find the deapth of map of images and from that generate the point cloud of scene.

Stereo-Dense-Reconstruction

Task

We are given 21 pairs of stereo images with calibration matrix and their Respective ground truth values, and also the baseline values from this data we have to reconstruct a 3d Point cloud representing all the points from the images.

Steps to get the Point clouds:

Get the Disparity Map from stereo image pair.

Math :
D = x1 - x2
Where x1 is the location of a point in the left image and x2 is the location of the point in the right Image.

Code :
Using the inbuilt function of Open CV, StereoSGBM_create using the tuning parameters of inspired by the blog post : http://timosam.com/python_opencv_depthimage.
And then using stereo.compute we calculate the disparity values.

Get the point cloud for a pair of images:

Math :
The 3d Point cloud of the images can be obtained by using these disparity values. The formula will be

Z = (bf)/(x1-x2)
X = (Zx) / f
Y = (Z * y) / f

Where:
b = baseline parameter provided in the question
f = Focal Length obtained from the K matrix
x = (x1+x2 /2)
y = (y1+y2 /2)

Code :
We do this operation using the Q matrix way, Were the Q matrix as defined in the Slides Q matrix. And Multiplied the Q matrix using Disparity_map with is [x,y,d,1]

Register the generated points and into world frame using the given ground truth values (poses.txt)

Math :
We have 3d point [wx,wy,w*z,w], and using the Projection matrices in ground truth we get the registered 3d point in the point cloud of a single world frame.

Code:
For each of the point in the point cloud multiply the point from the respective projection matrix and get and append these points into a single point cloud. And then visualize them.

OutPut: